MiniMax Releases M2.7, a Self-Evolving AI That Builds Its Own Training System
An AI That Trains Itself
Chinese AI startup MiniMax has released M2.7, a proprietary large language model that the company describes as "self-evolving." What makes it unique is that MiniMax used the model itself to build, monitor, and optimize its own reinforcement learning training infrastructure.
In other words, M2.7 helped create the very system that trained it — a concept that sounds like science fiction but is increasingly becoming a reality in AI development.
What Makes M2.7 Different
Most AI models are trained by human engineers who design the training pipeline, tune hyperparameters, and monitor progress. MiniMax took a different approach: they had an earlier version of the model help design and optimize the reinforcement learning harnesses used to train the next version.
The company claims this approach resulted in a model that performs 30-50% better on key benchmarks compared to what conventional training methods would achieve, while also being more efficient to train.
MiniMax's Rise in the AI Space
MiniMax has become one of the most exciting AI startups in China's crowded marketplace. The company has attracted significant investment and is positioning itself as a serious competitor to both Chinese giants like Alibaba and Western players like OpenAI and Anthropic.
The Bottom Line
A model that helps train its own successor is both impressive and a little unsettling. The "self-evolving" label is great marketing, but it also raises the question of where the feedback loop ends. If the AI is designing its own training, who is really in control of what it learns?